Base-Rate Betting on Polymarket: Start Every Market From the Reference Class
How to pick a reference class, convert a historical frequency into a market probability, adjust for specifics without storytelling, and size the bet when the base rate itself is uncertain.

Base-rate betting starts with a question that has nothing to do with the headline: how often does this class of event happen? Before you read the candidate's speech or the latest poll, count the reference class. The outside view is not a better forecast than the market. It is a check on whether your inside view is doing real work or just telling a story.
The method is simple to state and hard to apply. Define a class of comparable events, count how often the outcome occurred, adjust for the specific case, and compare the result with the executable price. The hard parts are choosing the class, avoiding one picked to fit the conclusion, and knowing how far the specific case should move the number.
Choosing a reference class
A useful reference class is narrow enough that the cases resemble the target and broad enough that there are enough cases to count. Twenty to one hundred cases is a practical range. Fewer and sampling error dominates; more and the class is probably too broad to be relevant.
For elections, the class could be incumbents seeking re-election in comparable districts, candidates leading by five points at this stage of a primary, or same-party holds after a retirement. Those are different classes with different base rates. The Federal Election Commission publishes official results, and the American Presidency Project maintains presidential election data going back to 1789.
For sports, the class could be home teams in game sevens or teams leading a best-of-seven series 2-0. Basketball Reference publishes the historical results. Match the format: a 2-0 lead in a best-of-seven is not a 2-0 lead in a best-of-five.
For economic data, the class could be CPI prints above consensus or recessions within 24 months of a yield-curve inversion. The Bureau of Labor Statistics and FRED publish the series. Match the definition: a recession has an official dating convention and a hot print does not.
Write the definition down before you look at the outcomes. If you try five classes and pick the one that matches your view, that is class shopping.
Outside view versus inside view
Kahneman and Tversky showed that people neglect base rates when given individuating information. A vivid description makes the specific story feel more informative than the statistical base rate, even when the description is unreliable. In markets, a debate performance feels like new information, so the base rate gets discarded.
Flyvbjerg's reference class forecasting makes the opposite case. Instead of asking a project team whether this project will finish on time, look at the distribution of actual completion times for similar projects. The outside view includes the failures that the inside view forgets.
The market price is also an outside view: it aggregates everyone's estimates. Your base rate has to be better than the market's prior, not just better than your gut. If your reference class says 35% and the market says 45%, you need a reason why the market's class is wrong, not just a feeling that 35% sounds right.
The arithmetic of adjustment
Start with a base rate b. Convert it to odds: b divided by (1 - b). For each specific factor, multiply the odds by a likelihood ratio. Then convert back to a probability. Keep the number of factors small, two or three, because each additional factor adds error and the factors are usually correlated.
Suppose your reference class is 40 comparable primaries, and candidates with the candidate's profile won 14 of them. The base rate is 35%. The odds are 0.35 divided by 0.65, or 0.538. The candidate leads by five points in the latest polls. In your historical sample, candidates with that lead won 60% of the time, so the likelihood ratio is 0.60 divided by 0.35, or 1.71. The adjusted odds are 0.538 times 1.71 = 0.920, and the probability is 0.920 divided by 1.920, or 47.9%.
That is the whole model. It is not precise, and the 35% base rate has its own error. With 40 cases, the standard error of the proportion is the square root of (0.35 times 0.65 divided by 40), or 7.5 percentage points. The base rate could be 27% or 43% on sampling error alone. Use a range, not a point.
The market is at 42 cents. Your adjusted estimate is 47.9%, so the gross edge is 5.9 cents. At 42 cents in a politics market, 1,000 shares pay a taker fee of 1,000 times 0.04 times 0.42 times 0.58 = $9.74, or 0.97 cents per share. The break-even probability is 42.97%. Your edge after the fee is 4.93 cents per share, or 11.7% of the 42-cent stake.
Sizing: the Kelly fraction for a binary buy is (q - p) divided by (1 - p). With q = 0.479 and p = 0.42, that is 0.059 divided by 0.58, or 10.2% of bankroll. Full Kelly is too aggressive when q is uncertain. Use a quarter, or 2.5%, and cap the event at 2%. The cap matters because several contracts in the same primary are one bet, not several.
Base rates for elections, sports and economic data
Election base rates are conditional. Incumbents win is different from incumbents win when the economy is strong. The narrower class is more relevant but has fewer cases; use it for the point estimate and the broad class as a sanity check.
Sports base rates are cleaner because the rules are stable. The main errors are format and era mismatches, so use the modern era and the exact format.
Economic data base rates are hardest because the process changes. A CPI surprise during an energy shock is not the same event as one in a stable year. The BLS series is revised, so use vintage data if you can.
Narrative traps
The first trap is this time is different. Sometimes it is, but the claim needs evidence, not a story. A candidate who raised more money or survived a scandal may genuinely be outside the class. The question is whether the change is measurable and already priced.
The second trap is the small-sample story. One comeback does not change a base rate. The third is selection: you remember the failures because they are interesting.
The fourth is conjunction. A candidate winning the nomination and then the general election is less likely than either event alone. Markets sometimes price the joint story as a single probability. If you can buy the two legs separately for less than the joint contract, the joint contract is overpriced, but the reverse is not automatically an arbitrage because the events are correlated.
How to execute
- Write the reference class in one sentence, including the time period and the rule set.
- Collect all the cases, not a convenient subset. Record the source and the date of the data.
- Compute the frequency and a confidence interval, not just a point estimate.
- List two or three specific adjustments with explicit likelihood ratios. If you cannot estimate a ratio, do not use the factor.
- Convert to a posterior probability and compare it with the executable ask, not the midpoint.
- Subtract the taker fee and the spread. If the edge is smaller than both, pass.
- Size with fractional Kelly and cap the total exposure to the event.
- Write down what evidence would change your mind before you enter.
Risks and limits
The main risk is the wrong reference class. A class that is too broad hides the relevant difference; a class that is too narrow has too few cases. The second risk is that the base rate is historical and the regime has changed. The third is that the market knows something your class does not, especially about turnout or a rule change.
Correlation is the fourth. Ten primary contracts in the same party are not ten independent bets. If you size each one as if it were independent, you will be far over your risk limit when the same swing moves all of them.
Finally, resolution risk sits underneath the probability. A contract that resolves on a different source or a different deadline is not the event your base rate describes. Read the rule before you count the cases.
A short checklist
- What is the reference class, exactly, and why is it the right one?
- How many cases, and what is the confidence interval?
- Which specific factors am I adjusting for, and what are their likelihood ratios?
- Is the market price already inside my range?
- What is the fee and spread hurdle at my size?
- How correlated is this bet with the rest of my book?
- What would make me abandon the base rate?
Base rates are not destiny. They are the starting line. The trade exists only when the specific case is different enough to move the probability and the market has not already made the same adjustment.
Sources
- 1.On the Psychology of Prediction— Psychological Review
- 2.Judgment under Uncertainty: Heuristics and Biases— Science
- 3.What You Should Know about Megaprojects and Why: An Overview— Project Management Journal
- 4.Election Results— Federal Election Commission
- 5.The American Presidency Project— University of California, Santa Barbara
- 6.Basketball Reference— Sports Reference